Your team's AI experiments are not an AI strategy

Walk through most organisations today and you will find AI everywhere and nowhere. Someone in marketing drafts with ChatGPT, an analyst has a clever Copilot habit, a partner quietly runs documents through Claude. It feels like momentum, and it is often mistaken for a strategy. A business with thirty people each using AI their own way does not have an AI strategy. It has thirty experiments and a rising sense that something is happening.

Experiments are good. They are how an organisation learns what works. But a pile of local efficiencies is not the same as a direction, and the gap between the two is where most of the value quietly leaks away.

Why busy is not the same as strategic

Scattered use has a ceiling, and the evidence is starting to show it. An MIT Media Lab report, The GenAI Divide: State of AI in Business 2025, found very high adoption sitting alongside very little measurable business return, a split it calls high adoption and low transformation. The figure deserves caution, since it is an industry report rather than peer-reviewed work and its headline number has been contested, but the shape of the finding matches what we see: plenty of activity, thin results.

There is a structural reason for this. Brynjolfsson, Rock and Syverson described a "productivity J-curve" in which the payoff from a general-purpose technology lags until organisations make the complementary changes around it: redesigned processes, shared practices, new responsibilities. Individual experiments cannot make those changes. They are too small and too private. Only a deliberate choice at the level of the business can, and that choice is what a strategy is. There is a competitive edge in this, too. When one firm turns scattered use into a few well-run, improving workflows while a rival leaves thirty private experiments to wither, the gap between them compounds quietly, year on year.

The AI strategy stack

A strategy does not have to be a long document. It has to answer seven questions in order, each building on the one above. We call it the AI strategy stack.

Business objective. What is AI in service of? Not "using AI", but the actual goal: more capacity without more hiring, faster turnaround clients notice, fewer errors in a risky process. If the top of the stack is vague, everything below it wanders.

Priority workflows. Which two or three pieces of real work matter most? A strategy is as much about what you are not doing as what you are. Spreading attention across ten workflows is how organisations end up with ten half-finished experiments.

Capabilities and tools. What does the chosen work actually need: a reusable assistant, a fixed automation, something more? Tools come after the work, not before. The order matters, and most organisations reverse it.

Data and knowledge. What does the AI need to see to be useful, and what must it never see? This is where most of the practical risk and most of the quality live. Good output depends on good context.

Human responsibilities. Who owns each AI-assisted workflow, who checks the output, and who is accountable when it goes wrong? The US National Institute of Standards and Technology builds its AI Risk Management Framework around a cross-cutting "Govern" function for exactly this reason: responsibility is set at the level of the organisation, not improvised by whoever happens to be using the tool.

Measures of value. How will you know it worked? One honest measure per workflow, agreed in advance, beats a dashboard of vanity metrics after the fact.

Review and learning. What is the rhythm for looking at what happened and deciding what changes? The tools move monthly, so a strategy that cannot update is out of date the day it is printed.

Scattered experiments are activity. A strategy is a decision.

What this looks like in practice

Consider an accountancy firm where a dozen people have their own AI habits. The experiments are genuinely useful, but no one can say whether they add up to anything. Running the stack changes the conversation. The objective becomes "release senior time from routine drafting." The priority workflows narrow to client letters and first-pass working papers. The tool is a shared, well-instructed assistant rather than everyone's personal account. The data rule names what client information must never be pasted in. Ownership sits with one partner. The measure is senior hours redeployed to advisory work. The review is monthly. None of that is grand, and all of it is a strategy. (An illustrative example, not a specific firm.)

The contrast with the starting point is the whole point. Before, the firm had activity it could not direct or defend. After, it has a direction it can resource, govern and improve. The government's own AI Opportunities Action Plan leans on the same instinct with its "scan, pilot, scale" sequencing: look before you leap, then leap deliberately.

The honest limits

Two cautions. First, a strategy can be overdone. A small business does not need a three-year AI transformation programme, and the seven layers should fit on a single page, not in a binder. The risk for most organisations is too little direction, but the opposite failure, planning instead of doing, is real and worth naming.

Second, none of this means stopping the experiments. They are the raw material a strategy is made from, and the teams already trying things are your best source of priority workflows. The job is not to shut the experimenting down. It is to harvest it into a shared decision, so the learning compounds instead of evaporating when one enthusiast moves on.

What to do next

Take one page. Write the business objective at the top. Name the two or three workflows that matter most, and cross out the rest for now. For each, note the tool it needs, the data it must and must not touch, who owns it, and the one measure that tells you it worked. Put a date in the diary to review it. That single page will do more for your AI than another round of scattered pilots.

The tool

To make that page easy to fill, we have built the One-Page AI Strategy Canvas: a single sheet with all seven layers, from business objective down to review and learning, with prompts under each so a leadership team can complete it together in one working session.

Download the One-Page AI Strategy Canvas (PDF)

Building that direction with a team is the work of an AI Reality Check Sprint: a short, focused engagement that turns scattered AI use into a small number of owned, measurable priorities. It builds on choosing the right first use case and on the difference between productivity and genuine business value.

Sources and further reading

• MIT Media Lab (Project NANDA), The GenAI Divide: State of AI in Business 2025. Industry report, not peer-reviewed; the 95% figure has been contested. Directional evidence of high adoption with low transformation.
• Brynjolfsson, Rock and Syverson, The Productivity J-Curve, NBER Working Paper 25148, 2018. Independent. Source for the lag between adoption and value, driven by organisation-level complementary change.
• NIST AI Risk Management Framework. Independent US standards body. Source for the cross-cutting "Govern" function.
• UK Government, AI Opportunities Action Plan, January 2025. Source for the "scan, pilot, scale" sequencing of adoption.